AbstractMany neural networks have been studied, most of them using computer simulations, as a source of new computational paradigms. Since most implementations of neural networks require many elemental components, it is necessary to use a fault‐tolerance mechanism with them. This paper examines a hierarchical neural network in which recovery from a functional breakdown of the whole network (due to breakdown of its parts) can be performed by relearning. The repeated relearning significantly improves the time to failure of the whole network after a partial breakdown. © 2002 Wiley Periodicals, Inc. Electron Comm Jpn Pt 2, 85(11): 81–88, 2002; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjb.10097